Summary of Discriminator Soft Actor Critic Without Extrinsic Rewards, by Daichi Nishio et al.
Discriminator Soft Actor Critic without Extrinsic Rewards
by Daichi Nishio, Daiki Kuyoshi, Toi Tsuneda, Satoshi Yamane
First submitted to arxiv on: 19 Jan 2020
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Machine Learning (stat.ML)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper addresses the challenge of imitating expert behavior in unknown states from limited data. Supervised methods like Behavioral Cloning don’t require sampling data, but often struggle with distribution shift. Reinforcement learning-based approaches like inverse reinforcement learning and generative adversarial imitation learning (GAIL) can learn from few expert examples, yet typically require environment interaction. Soft Q imitation learning combines Behavioral Cloning and soft Q-learning for efficient learning, while our proposed Discriminator Soft Actor Critic (DSAC) enhances robustness to distribution shift by using an adversarial inverse reinforcement learning reward function. We evaluate DSAC on PyBullet environments with only four expert trajectories. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about teaching a machine to mimic what experts do in situations we don’t know much about, using very little data. The usual ways of doing this, like copying the experts’ behavior, aren’t perfect because they can’t handle big changes in the situation. Another approach is to use reinforcement learning, which can learn from just a few expert examples, but it usually needs to try out different actions itself. A new method combines these ideas and does well with very little data. To make this even better, we came up with a way to add some extra protection against changes in the situation. We tested this on simple computer simulations using only four expert examples. |
Keywords
* Artificial intelligence * Reinforcement learning * Supervised
